H Company (Hcompany) published NeoMME on the Hugging Face blog on September 3, 2026, introducing a family of 260M and 800M multilingual multimodal encoders. According to the post, NeoMME does not use a separate pretrained vision tower or a causal language model. Instead, a single bidirectional Transformer processes both text tokens and raw image patches, trained from scratch with a masked discrete-diffusion objective. Each model processes about 524 billion packed input tokens, including 290 billion from text-only examples, which the authors note is small compared with ModernBERT's 2 trillion token budget; they used the NorMuon optimizer to improve data efficiency. The team fine-tuned the models for visual document retrieval using ColPali's page-image approach, producing NeoMME-Retriever, which returns dense and late-interaction embeddings in one forward pass. The post states both sizes lie on the ViDoRe v3 Pareto frontier for nDCG@10 versus model size. At a 2048×2048 image input on an NVIDIA L40S GPU, the 260M model encodes about 51 pages per second, roughly twice ColModernVBERT's throughput. Hierarchical token pooling and asymmetric quantization reduce late-interaction index storage from roughly 1.5 MB to 6 kB per page (255× smaller) while retaining more than 95% of baseline nDCG@10. All checkpoints are released under the Apache 2.0 license.
- Two model sizes: 260M and 800M parameters, released under Apache 2.0
- Single bidirectional Transformer handles both text and image patches, trained from scratch
- 260M model encodes ~51 pages/second at 2048×2048 on an NVIDIA L40S GPU
- Index storage cut from ~1.5 MB to 6 kB per page (255× smaller) retaining >95% of nDCG@10
What it means for you
This is a new set of open-source models built to search through document images — think scanned PDFs, invoices, or report pages — without first converting them to text. Because it treats each page as a picture, it keeps layout, tables and charts that regular text extraction loses. The headline practical win is efficiency: it stores search indexes far more compactly and runs faster, which lowers the cost of building document-search systems.
Try this
If you already run a document-retrieval or RAG setup, have a developer benchmark NeoMME-Retriever against your current OCR-based pipeline on a sample of your own documents, since it's free under Apache 2.0 and available in Hugging Face Transformers.
Who should care
Technical teams and developers building document search, retrieval, or visual RAG systems, especially those handling multilingual documents or image-heavy PDFs where OCR struggles.
Skip this if
You are a non-technical reader or small business without an in-house AI/dev capability — this is a building block for engineers, not an off-the-shelf product you would use directly.
Sources: Hugging Face — read the original